acceptodds
Under review as a conference paper at ICLR 2027

Sparse cubical complexes for efficient topology-preservation in image data

Abstract

Persistent homology (PH) is a frequently used tool for extracting and preserving topological information from image data, particularly in image segmentation, where preservation of topological structures is important. However, despite its general applicability across dimensionality, domains, and target structures, the runtime cost of PH-based methods often makes their practical use infeasible. In this work, we argue that this runtime cost is largely driven by processing information that is unimportant for downstream application (e.g. as optimization objective). We propose sparse cubical filtrations as an alternative foundation for PH computation, reducing subsequent computational costs by factors of up to 100 on real datasets. We show close agreement with the optimization signal of the dense counterpart and empirically evaluate our solution's effectiveness as an optimization objective in realistic training regimes where other PH-based objectives can practically not operate (i.e., 3D data with large patch sizes). We show how our solution improves topological accuracy by up to 80% across six diverse datasets while maintaining pixel- and region-based accuracy.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
persistent homology, topology, image segmentation, topology-preserving image segmentation, loss function

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_SQckOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes sparse cubical complexes to accelerate persistent homology (PH) computations in topology-aware image segmentation. Following Betti Matching, the proposed sparseBM makes loss computation up to about 100 times faster. It is evaluated on practical 3D segmentation datasets.

Critical strengths

The method makes PH-based losses fast enough to be practical for 3D segmentation. Loss computation becomes up to about 100 times faster, and the training overhead on 3D data is small, so a PH-based loss can be used with realistic patch sizes, where the original Betti Matching is infeasible.

Critical weaknesses
  1. The relation to Betti Matching and the method details are not clearly presented

The method follows "the principles of Betti Matching", but the method section does not explain the procedure clearly (Betti Matching is described only briefly in the related work). The constructions used in the experiments must also be pieced together (S, C=min(L, P), the convention that filtration values are one minus the predicted probability, etc). Moreover, discarding cells above a threshold is a standard technique (e.g., Ripser and Cubical Ripser), so presenting the contribution as "sparse cubical complexes that omit high-valued regions" somewhat overstates its novelty, although retaining the original filtration values on a common comparison-based subcomplex is a sensible refinement.

  1. The evaluation is not very fair

Since sparseBM optimizes a Betti Matching objective, using the Betti Matching error as the main topological metric favors the proposed method, and the claim that it "substantially reduce topological errors" rests mainly on this metric. No meaningful changes in clDice, VOI, or NSD are observed. Also, the model selection is aligned with the same metric. For every method, the loss weight is chosen by the lowest validation Betti Matching error.

TopoMortar (Valverde et al., BMVC 2025, https://bmvc2025.bmva.org/proceedings/596/) pointed out pitfalls in evaluating topology losses. Data augmentation, which is under-reported in the literature, strongly affects topology accuracy and can make a Dice+CE baseline surpass most topology losses, and with few seeds any loss can be made to appear best. The submission states that it follows nnU-Net principles but without specifying the augmentations, and its results and paired t-tests rely on three seeds.

Questions for authors
  1. Could the authors write out the sparseBM loss explicitly, including its combination with Dice+CE and the selected weight for each dataset?
  2. How does sparseBM compare with the baselines on topological metrics it does not directly optimize?
  3. Can the main results be reported with more seeds?
11 comments
AuthorsOpenReview ↗

More detailed description of procedure and parameters

We thank the reviewer for the insightful comments and the helpful ideas on how to improve the paper further. We provide detailed answers below. Each raised topic is structured into three parts: a brief summary of the reviewer's feedback, a TL;DR reply, and a more detailed reply with evidence and references to the updated manuscript. We look forward to receiving further feedback on the explanations provided and the updated manuscript.

R SQck: The relation to Betti Matching and the method details are not clearly presented. Write out the sparseBM loss explicitly, including its combination with Dice+CE and the selected weight for each dataset.

TLDR. We added a corresponding description to the method section in the manuscript, in which we explain the loss computation procedure in detail. In addition to Figure 7 in the supplement, we added Table 5, listing the weights used for each dataset. The weights for all methods are as follows:

DatasetclDice (α)Skel. RecallWarpingDMTTopographSparse BM (ours)
ATM'260.10.0635e-3––0.02
NISB-B0.750.256.7e-7––1.7e-4
BraTS-METS0.0250.258.2e-4––0.19
MMWHS0.750.50.12––0.032
FIVES (2D)0.2512.1e-53600.0194e-3
ACDC (2D)0.50.250.058780.140.27

For all methods, we conducted hyperparameter searches to determine the optimal loss weight. Contrary to sparseBM (Figure 7), we did not find a rule to "predict" the best weights for the comparison methods. Especially fixed weights from the respective papers turned out to be suboptimal on many datasets.

AuthorsOpenReview ↗

BM as primary metric

R SQck: The paper uses BM error as the primary metric, and it is related to the proposed training objective. Metrics that are not directly optimized should also be used.

TLDR. We now evaluated all methods on two additional metrics that measure pure topology and are not directly related to our method. Furthermore, we added three downstream evaluations on metrics connected to the datasets' use cases. Our method's improvements persist in all these additional measurements. Many other metrics (VOI, NSD) are unsuitable for measuring pure topology (Berger et al., IPMI 2025), and we provide simple explanations and examples below.

We believe that BM error is the most principled metric for capturing pure topological accuracy, accounting for spatial correspondence of topological features. To further assess topological accuracy, we added two additional metrics that are not directly optimized by our method: the global difference in the number of topological features (Betti-number error; Hu et al., NeurIPS 2019) and the DIU (discrepancy of intersection and union; Lux et al., ICLR 2025). We added these metrics to the main results table and show significant improvements across both. Due to space limitations in the main table, we moved NSD and VOI to the appendix. We are open to including the results the reviewers deem most appropriate in the final main tables and look forward to their feedback on this.

We agree that alternative metrics (VOI, NSD, clDice) are informative for certain downstream tasks and that they often correlate with topological correctness. However, none of them measure topology itself. Each is a geometric or information-theoretic surrogate that also responds to region size, boundary position, or overlap, and each can miss topological errors (Berger et al., IPMI 2025, "Pitfalls of topology-aware image segmentation"). The mentioned pure topology metrics (BM error, Betti number error, DIU) are not affected by region sizes or boundary displacements. We provide minimal 2D examples that showcase the differences at the end of this response.

Furthermore, we present three examples to demonstrate that the measured topological improvements can yield gains on the downstream task. For a more detailed description of these results, we refer readers to the new Appendix A.5.

Improvement in ejection fraction error and closed ring rate on ACDC. In every short-axis slice, the myocardium forms a closed ring around the left-ventricular cavity. If the ring is broken, the cavity and thus the ejection fraction (EF), the key clinical measure of cardiac function, cannot be computed correctly. sparseBM reduces the share of broken rings by 43% compared to the best baseline (2.4% vs. 4.2%) and lowers the EF error from at least 3.31 to 2.75 percentage points, both significantly better than all baselines.

Improvement in lesion detection on BraTS-METS. For brain metastases, the number of lesions is clinically relevant, e.g., for treatment planning. Under the official BraTS lesion-wise evaluation, sparseBM improves the lesion detection F1 score from at most 0.744 to 0.778, significantly better than all baselines, and achieves the best lesion-wise Dice and HD95.

Improvements on tree length detected (TLD) and branch detection (BD) in airway tree segmentation. As presented in Appendix A4.2, we applied our method as a post-processing step on airway tree segmentations. Across all base segmentation networks, the post-processing improved TLD and BD. Although we did not use any additional optimizations (e.g., TTA, ensemble methods, etc.), our method achieved a top ranking in this year's ATM challenge, underscoring its real-world effectiveness.

Despite these gains, we note that our method focuses on enhancing general topological accuracy. The question of whether this corresponds to improved downstream performance for domain-specific tasks is not strictly in the scope of this paper. Notably, Betti matching, and thus sparseBM, allows applying the loss to specific dimensions only, e.g., to connected components for lesion detection or to loops for closed myocardial rings, which can be relevant for downstream tasks. However, since we aim to show the general working mechanism, we do not adjust the loss to favor specific downstream evaluations/tasks.

AuthorsOpenReview ↗

Examples for conflating factors of NSD, clDice, and VOI metric

Minimal 2D examples showcasing the conflation of region sizes and boundary location in NSD, VOI, and clDice (foreground 4-connected, background 8-connected, NSD tolerance τ = 1 px):

(a) break   (b) hole    (c) shifted (d) dilated (e) dent
.........   .........   .........   .........   .........
.........   .#######.   .........   .+++++++.   .###o###.
..##o##..   .###o###.   ..ooooo..   .+#####+.   .#######.
..##o##..   .#######.   ..ooooo..   .+#####+.   .#######.
..##o##..   .#######.   ..#####..   .+#####+.   .#######.
.........   .#######.   ..+++++..   .+++++++.   .#######.
.........   .........   ..+++++..   .........   .........
.........               .........   .........

# prediction and ground truth, o ground truth only, + prediction only, . background

Betti err.BM err.DIUVOI ↓NSD ↑clDice ↑
(a) break1110.561.000.86
(b) hole1110.200.981.00
(c) shifted0001.440.580.00
(d) dilated0001.220.881.00
(e) dent0000.201.001.00

The three topological metrics are 1 exactly for the two topological errors (a, b) and 0 otherwise. Each surrogate scores a topologically correct prediction worse than a topologically wrong one.

Reviewer_HbXBOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

In this work, the authors tackle computing persistent homology (PH) in an efficient manner for computer vision (CV) tasks. Specifically, the use of PH is in CV is often limited when size of image/volume is large because PH is expensive to compute. Here the authors leverage the fact that most areas are homogenous. Hence they construct sparse complexes by omitting cells representing confident background regions and use them as the basis of their optimization objective, sparseBM

Critical strengths
  • This is an important contribution in making PH adoption widespread for large-sized images and volumes.
  • The approach is well-motivated, and the assumptions are reasonable: (1) most datasets and network predictions are sparse in the foreground, and (2) confident background regions contain no information that is necessary for computing topology-preserving loss functions.
  • I appreciate authors' extensive experiments (6 datasets across 2D and 3D, and comparing against several baselines)
  • Table 2, it is great to see comparison with dense BM (where feasible), both improvement in metrics and reduction in train time
Critical weaknesses
  • L19 says "reducing subsequent computational costs by factors of up to 100" but this does not seem to be the case within the paper (Fig 4 & Fig10).
  • Figure 7 says "ACDC 3D" but the experiments are 2D
Questions for authors

see weakness

1 comment

We thank the reviewer for the positive review and for pointing that our method is well motivated, based on reasonable assumptions, and extensively evaluated. We reply to the points raised below.

R HbXB: Wrong caption in Figure 7.

We've corrected the error in the figure.

R HbXB: The abstract claims a reduction in computational complexity up to a factor of 100, but this is not supported by Figure 4 and Figure 10 (now 11).

The mentioned factor refers to Figure 1, in which we compared the runtime up to a patch size of 1923192^3, yielding a speedup of 107107 compared to Betti Matching. In our main experiments, we mostly used a patch size of 128128 to accommodate our GPU capacities. The speedup increases with patch size, and we scaled up to 1923192^3 because we believe this is a realistic patch size for training segmentation networks. If the reviewer finds the reference to Figure 1 misleading, we are happy to revise the abstract to report the speedups from Figure 4 instead.

Reviewer_3XzQOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper targets the high computational cost of persistent-homology-based losses for image segmentation, especially in 3D large-patch training. The authors propose sparse cubical filtrations: rather than constructing the full dense cubical complex, they retain a comparison-filtration subcomplex and implicitly represent omitted high-confidence background regions. This leads to sparseBM, a sparse variant of Betti Matching.

The validation combines theoretical comparison to the dense filtration, runtime measurements, segmentation experiments on six 2D/3D datasets, tau and loss-weight ablations, and a reduced-patch comparison to dense Betti Matching. The empirical claim is that sparseBM preserves much of the dense optimization signal while reducing PH computation by large factors and improving Betti Matching error with little Dice degradation.

Critical strengths
  1. The paper addresses a real bottleneck in PH-based segmentation losses. The reported speedups are large enough to change practical feasibility, especially for 3D large-patch training where dense PH methods are otherwise not usable.

  2. The empirical study is unusually broad for this topic: six datasets, 2D and 3D settings, multiple topology-aware baselines, runtime analysis, tau ablations, and a reduced setting comparing against dense Betti Matching. This breadth supports the claim that the method is not just a small synthetic demonstration.

Critical weaknesses
  1. The primary success metric, BM error, is very close to the proposed training objective. This makes the headline improvement less independently convincing. VOI, NSD, clDice, Dice, and qualitative downstream relevance show weaker or mixed changes, so the paper needs stronger evidence that BM-error reduction corresponds to practically meaningful topology correction.

  2. The theoretical discussion correctly states that sparse and dense Betti matchings are not guaranteed to be pairwise equal above tau, but the main narrative sometimes reads as if sparseBM is a near drop-in replacement for dense BM. The method's failure modes need to be made more explicit: dense foregrounds, uncertain backgrounds, early training predictions, non-sparse targets, and cases where omitted terminal-band features matter. In the current version, this leaves the paper short of the evidentiary bar for acceptance despite the useful idea.

Questions for authors
  1. Can the authors provide qualitative and task-specific evidence that BM-error reductions correspond to meaningful corrections, not only improvements under a loss-aligned metric?

  2. How sensitive is the method to the fixed choice tau = 0.8 during early training, when model predictions may not yet have large confident-background regions? Would an adaptive tau or curriculum improve robustness?

  3. For datasets with large foreground fractions or dense topological structures, at what retained fraction does sparseBM lose its practical advantage over dense BM?

  4. In the reduced-patch dense-BM comparison, are denseBM and sparseBM implemented with exactly the same optimized backend and hardware path? Please clarify to make the runtime comparison fully reproducible.

6 comments
AuthorsOpenReview ↗

Using Betti Matching error as primary metric

We thank the reviewer for the detailed feedback and suggestions on how to further improve this work. We provide detailed answers below. Each raised topic is structured into three parts: a brief summary of the reviewer's feedback, a TL;DR reply, and a more detailed reply with evidence and references to the updated manuscript. We look forward to the reviewer's thoughts on our responses and the revised manuscript

R 3XzQ: The primary success metric, BM error, is very close to the training objective. Can the authors provide qualitative and task-specific evidence that BM-error reductions correspond to meaningful corrections, not only improvements under a loss-aligned metric?

TLDR. We added qualitative results to the manuscript and evaluated our method in three task-specific settings. Furthermore, we evaluated all methods on two additional topological metrics that are not directly related to our method. Our method’s improvements persist in all these additional measurements.

Our motivation for using BM error as our primary metric is that we believe it is the most principled metric for capturing pure topological accuracy, accounting for spatial correspondence of topological features. To further assess topological accuracy, we now added two additional metrics that are not directly optimized by our method: the global difference in the number of topological features (Betti-number error; Hu et al., NeurIPS 2019) and the DIU (discrepancy of intersection and union; Lux et al., ICLR 2025). We added these metrics to the main results table and show significant improvements across both. Due to space limitations in the main table, we moved NSD and VOI to the appendix. We are open to including the results the reviewers deem most appropriate in the final main results and look forward to their feedback on this.

Additionally, we present three examples to demonstrate that the measured topological improvements can yield gains on the downstream task. For a more detailed description of these results, we refer readers to the new Appendix A.5.

Improvement in ejection fraction error and closed ring rate on ACDC. In every short-axis slice, the myocardium forms a closed ring around the left-ventricular cavity. If the ring is broken, the cavity and thus the ejection fraction (EF), the key clinical measure of cardiac function, cannot be computed correctly. sparseBM reduces the share of broken rings by 43% compared to the best baseline (2.4% vs. 4.2%) and lowers the EF error from at least 3.31 to 2.75 percentage points, both significantly better than all baselines.

Improvement in lesion detection on BraTS-METS. For brain metastases, the number of lesions is clinically relevant, e.g., for treatment planning. Under the official BraTS lesion-wise evaluation, sparseBM improves the lesion detection F1 score from at most 0.744 to 0.778, significantly better than all baselines, and achieves the best lesion-wise Dice and HD95.

Improvements on tree length detected (TLD) and branch detection (BD) in airway tree segmentation. As presented in Appendix A4.2, we applied our method as a post-processing step on airway tree segmentations. Across all base segmentation networks, the post-processing improved TLD and BD. Although we did not use any additional optimizations (e.g., TTA, ensemble methods, etc.), our method resulted in a top ranking in this year's ATM challenge, underscoring its real-world effectiveness.

Despite these gains, we note that our method focuses on enhancing general topological accuracy. The question of whether this corresponds to improved downstream performance for domain-specific tasks is not strictly in the scope of this paper. Notably, Betti matching, and thus sparseBM, allows applying the loss to specific dimensions only, e.g., to connected components for lesion detection or to loops for closed myocardial rings, which can be relevant for downstream tasks. However, since we aim to show the general working mechanism, we do not adjust the loss to favor specific downstream evaluations/tasks.

Furthermore, we now include qualitative results in the new Appendix A.6.

AuthorsOpenReview ↗

Sensitivity to tau during early training

R 3XzQ: How sensitive is the method to the fixed choice tau = 0.8 during early training, when model predictions may not yet have large confident-background regions? Would an adaptive tau or curriculum improve robustness?

TLDR. Model predictions obtain large confident background regions very fast. After a warm-up period, our method is robust to different choices of tau in the range 0.8−0.990.8-0.99. Without a warm-up period, a curriculum-based tau can improve robustness.

We experimented with varying tau's combined with different warm-up lengths (i.e., number of epochs before enabling sparseBM) and observed no consistent impact on performance. For each warm-up length, the choice of tau did not consistently affect the final performance or runtime. The behavior reported in Table 3 and Figure 8 is largely independent of the warm-up lengths.

However, having no warm-up, i.e., enabling the loss on a randomly initialized network, destabilizes training, and performance can often not be fully recovered. This is caused by the many topological features of random predictions and the resulting large sparseBM gradient.

In the table below (and Supplement Figure 13), we measure the retained fraction and the ratio of sparseBM's gradient norm during very early training (as an offline measurement, i.e., without actually using sparseBM's gradient for training). It shows that the retained fraction drops to a few percent within 10 iterations and is close to its value at our switch-on point from epoch 2 on. Therefore, the sparsity, and thus the efficiency, of sparseBM is restored almost immediately. However, the gradient ratio between sparseBM and Dice+CE is still much higher than at our switch-on point, and it falls to within 2× from epoch 3 on. A network with random weights creates many topological features, resulting in a large gradient that dominates early training.

ATM'26FIVES
Training stage (τ = 0.8)RetainedGrad. ratioRetainedGrad. ratio
Iteration 0 (random initialization)66 %–75 %–
Iteration 536 %–34 %–
Iteration 103.1 %217×12.0 %58×
Epoch 1 (mean)3.7 %17.6×15.0 %12.0×
Epoch 2 (mean)1.7 %2.7×9.8 %1.6×
Epoch 3 (mean)1.5 %2.0×8.9 %1.0×
Epoch 5 (mean)1.4 %1.4×8.4 %0.9×
Epoch 10 (mean)1.3 %1.7×8.2 %0.9×
After warm-up (sparseBM is switched on here)1.3 %1×8.0 %1×
Label foreground (reference)1.2 %7.6 %

Dynamically adjusting tau can (at least partially) mitigate this problem. We ran preliminary experiments (single seed, only ACDC dataset) and evaluated three different curriculum-based approaches (see Figure 14), in which the loss is used without warm-up, and tau is gradually increased as training progresses. We added details of these experiments to Appendix A.7. We show that a gradually increasing tau reduces the initial gradient ratio from up to 400× to ~16× of its value at our switch-on point (Figure 15). This results in increased test performance at the end of training compared to a fixed tau without warm-up. However, a sufficient warm-up achieves the same (or better) effect more simply, which is why we recommend it.

Warm-up (epochs)τDice ↑BM err. ↓
Dice+CE (no sparseBM)–0.89700.255
00.80.87790.145
00.01 → 0.8 over 1 epoch0.89080.121
00.01 → 0.8 over 10 epochs0.88710.122
00.01 → 0.8 over 100 epochs0.89250.141
10.80.88980.167
100.80.88860.144
100 (default)0.80.89300.109

Conclusion. A dynamic or curriculum-based tau value can increase robustness when no warm-up period is applied. However, we generally recommend using sparseBM (and other non-normalized PH-based losses) after a warm-up phase because we did not observe any benefit when using it early on. After a warm-up phase, we observe high robustness to the choice of tau, in both runtime and performance (see Table 3 and Figure 8).

AuthorsOpenReview ↗

Break-even point with increasing retained fraction

R 3XzQ: At what retained fraction does sparseBM lose its practical advantage over dense BM?

TLDR. sparseBM has a runtime advantage over dense BM up to a retained fraction of around 60% (compared to the 1% to 8% encountered in our datasets).

We have created a synthetic dataset with random labels and varying foreground (i.e., also retained) fractions. The break-even point is at around 60%. We added the experiment to the supplement in Figure 9 and mention the break-even point in the limitations section.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the discussion period

Dear reviewers and AC,

Thank you for your time and valuable feedback. We genuinely believe that the paper has improved as a result of the discussion period. Below is a brief summary of the discussion and the resulting changes.

All three reviewers agree that sparseBM addresses a real bottleneck. It makes PH-based losses practical for 3D segmentation at realistic patch sizes, where dense methods cannot be used (3XzQ, HbXB, SQck). The reviewers also found the approach well-motivated and its assumptions reasonable (HbXB). They valued the breadth of the evaluation, which 3XzQ called “unusually broad for this topic”.

Main discussion points and changes

  • BM error as primary metric (3XzQ, SQck). We added two additional topological metrics, which both confirm our ranking. We also added exploratory downstream evaluations. Nonetheless, we chose BM error because it is the most principled measure of pure topological accuracy. Therefore, we argue that sparseBM is a metric-aligned loss rather than BM being a loss-aligned metric.
  • Statistical evidence (SQck). We now use five seeds instead of three, specify all augmentations (Table 6), and apply Holm’s correction for our main results.
  • Novelty relative to filtration truncation (SQck). Section 3 and the new App. A.8 now explain why truncating a single filtration does not work for matching. We have also added empirical evidence for that.
  • Robustness and limitations (3XzQ).
    • New analysis of τ and warm-up, including a curriculum τ (App. A.7).
    • Break-even point is at about 60% retained fraction, versus 1–8% in our datasets (Fig. 9).
    • The limitations section now states the failure modes explicitly.
    • We confirmed that dense and sparse BM were run on the same backend and hardware.
  • Clarifications (SQck). We wrote out the loss explicitly (Eqs. 6–7) and added the weights for each dataset (Table 5).

Thank you again for your time and effort.

Best regards, The authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.